This repository contains the official implementation of
Taehwan Lee and Sung Whan Yoon, Rethinking the Flat Minima Searching in Federated Learning, International Conference on Machine Learning (ICML) 2024.
We refer to the FedLab for creating our project.
- install conda environment (preferred):
conda env create -f environment.yaml
- The code runs with WANDB. For setting up your profile, we refer you to the quickstart documentation.
- WANDB MODE is set to "online" by default.
- If you set
args.wandb_project_name
asdebug
, WANDB will be 'disabled'. - You also can switch to "offline" Here.
- Overview: Image Dataset based on CIFAR and Federated Vision Datasets
- We store the json file that distribute the images with Dirichlet's distribution in
tools/json_data
. - 100 users have 500 images each. Different
$\alpha$ value is possible in Dirichlet's distribution.
An example command can be found in tools/experiments
cd tools/experiments
chmod +x cifar.sh
./cifar.sh
@inproceedings{
leerethinking,
title={Rethinking the Flat Minima Searching in Federated Learning},
author={Lee, Taehwan and Yoon, Sung Whan},
booktitle={Forty-first International Conference on Machine Learning}
}